Research On Asphalt Pavement Crack Identification Method Based On Semantic Segmentation Network

As an important part of the modern infrastructure system,highways have greatly promoted the economic and social development and facilitated people’s travel.However,with the increase of road age,various diseases will appear on the road surface.As an early manifestation of various diseases,efficient and accurate identification of road surface cracks and their maintenance can avoid further development of cracks and save maintenance costs.This paper takes the asphalt pavement crack images collected by the road detection vehicle as the research object,based on the research method of deep learning,aims to design an efficient and accurate semantic segmentation network for early crack identification and realize preventive maintenance of asphalt pavement.The main research contents of this paper are as follows:(1)Aiming at the problem that the width characteristics of cracks in existing public crack datasets are more obvious and the shooting angles of images are different,this paper first uses a road detection vehicle for road image acquisition,selects crack pictures and makes experimental datasets.The width characteristics of this dataset are thin,the shooting angles are unified and the image background is rich;secondly,data enhancement and preprocessing are carried out on the experimental dataset to increase the number of experimental datasets and meet the network input requirements.(2)Crack recognition based on semantic segmentation network.Firstly,by comparing the application effects of four semantic segmentation networks on experimental datasets,this paper determines the basic network UNet;secondly,based on UNet,a lightweight segmentation network EDSR＿UNet is proposed.This network combines improved depth separable residual structure.Compared with UNet,this network reduces the calculation and parameter requirements of the network to 1/6 of the original under the condition that the recognition accuracy is only reduced by 0.2%;finally,different initial learning rates are compared for their impact on network convergence speed.The experimental results show that when the initial learning rate is 0.01 in 500 rounds of experiments,the effect is best,and recall rate,average intersection ratio and F1 score are highest,which are 83.83%,68.59% and 81.37%,respectively.(3)Aiming at the problem that crack recognition effect in complex road surface images with interference factors such as lane lines and linear oil stains is poor,a residual dense UNet(RD-PSA＿UNet)with pyramid squeezing attention is proposed.Firstly,the convolution layers in UNet encoder part are replaced by residual dense blocks to strengthen information interaction between layers through feature reuse;secondly,pyramid squeezing attention module is integrated into residual dense block to make each channel information output carry attention weight and enhance contribution of key information.Comparative experiments and ablation experiments show that RDPSA＿UNet has improved comprehensive detection effect.Compared with UNet,F1 score of this network is increased by 1.06%,average intersection ratio is increased by 1.52%,and precision is increased by 2.92%.(4)Aiming at the problem that direct input of large-size images into network may cause memory overflow,this paper proposes a method of cropping large-size images before recognition and then merging recognition results into original image prediction graph.Moreover,a crack recognition system based on road detection vehicle for collecting road surface images is designed and implemented for this application scenario.